plawanrath/StableHLO-OutOfGrammar-25

StableHLO-OutOfGrammar-25是一个包含25个手工编写的自然语言到StableHLO对的数据集，这些对使用了超出10-op语法范围的运算符。该数据集是六个NL→MLIR基准测试之一，用于评估模式派生的约束解码在MLIR中的跨方言泛化能力。数据集格式为每行一个JSON记录，包含方言、ID、MLIR、自然语言、操作和来源等字段。所有参考MLIR程序在发布时都经过验证器验证，且为单人手工编写，不包含众包或LLM生成的内容。数据集主要用于测试，不适合用于微调以避免污染未来评估。

StableHLO-OutOfGrammar-25 is a dataset of 25 hand-authored NL→StableHLO pairs using ops outside the 10-op grammar scope. It is one of six NL→MLIR benchmarks for evaluating schema-derived constrained decoding for cross-dialect generalization in MLIR. The dataset format is one JSON record per line, with fields including dialect, id, mlir, nl, op, and source. All reference MLIR programs are verifier-clean at release, hand-authored by a single author without crowdsourcing or LLM-generated content. The dataset is test-only and not suitable for fine-tuning to avoid contaminating future evaluations.